利用深度學習與水下影像研發白蝦成長及行為模式的物聯網監測系統

Abstract

中南美白對蝦是目前全球養蝦產業最重要的物種，即時觀測白蝦的行為模式有助於提高飼料效率和降低成本，傳統上白蝦的觀測是透過傘網，此種做法費時費力，且會干擾白蝦的活動。本研究研發白蝦物聯網監測系統，包含白蝦即時影像觀測、白蝦偵測模型、投餌參數分析和自動投餌等功能。本研究研發多種數位和類比式水下影像系統及附件，搭配樹苺派控制和紅外線照明，可配合三種養殖池型，提供串流影像的即時觀測及備份，能即時觀測六種水色的白蝦影像，包括蝦苗或夜間影像。本研究利用深度學習研發白蝦偵測模型，將白蝦影像分成蝦體跟尾扇，使用卷積類神經網絡的模型YOLOv4來偵測白蝦的位置，利用影像處理將白蝦與背景分離，計算白蝦的體長、拖糞率和出現數，合稱投餌參數，每間隔一周對白蝦影像做24小時連續性的統計分析，共歷時兩個半月，以歸納出白蝦的生長及行為模式。本研究在彰化、高雄及屏東三處養蝦場進行監測系統的驗證，結果顯示，YOLOv4模型辨識白蝦的平均精準度（mAP）為0.76，出現數的精準度為0.85，拖糞率的精準度為0.69，體長的精準度為0.94。體長與時間呈線性變化，決定係數(R square)為0.98，適用於3公分以上的白蝦。影像系統平台高度會影響蝦子出現的機會，出現數在投餌後會有明顯下降再上升的趨勢，顯示白蝦被飼料吸引離開平台上方的程度，有助於判斷白蝦食慾，調整投餌量及投餌時機，拖糞率在前半段的變化趨勢與出現數相似，後半段則不顯著。本研究研發的水道池專用的自動投餌機，可雲端控制投餌，讓夜間投餌成為可行，在屏東完成全自動投餌實驗。本研究所提出的白蝦物聯網監測系統克服了池塘養殖中白蝦不易觀測的瓶頸，為白蝦生長及行為模式提供重要資訊，並有助於投餌控制的自動化。

Penaeus vannamei is currently the most important species in the global shrimp farming industry. Real-time observation of the behavior pattern of white shrimp can help improve feed efficiency and reduce costs. Traditionally, white shrimp is observed through feeding trays, which is time-consuming and labor-intensive and interferes with the activity of white shrimp. This study develops the white shrimp IoT monitoring system, which includes the functions of white shrimp real-time image observation, white shrimp detection model, feeding parameter analysis and automatic feeding. In this study, a variety of digital and analog underwater imaging systems and accessories have been developed. With Raspberry Pi control and infrared lighting, they can be matched with three types of aquaculture ponds, providing real-time observation and backup of streaming images, and can instantly observe six colors of white shrimp, including shrimp fry or nighttime image. This study uses deep learning to develop a white shrimp detection model, divides the white shrimp image into shrimp body and tail fan, uses the convolutional neural network model YOLOv4 to detect the location of the white shrimp, and uses image processing to separate the white shrimp from the background. Calculate the body length, dripping rate and appearing number of the white shrimp, which are collectively referred to as the feeding parameters. Statistical analysis of the white shrimp image for 24 hours is performed every one week for two and a half months to summarize the growth and behavior of the white shrimp. In this study, the monitoring system was verified in three shrimp farms in Changhua, Kaohsiung and Pingtung. The results showed that the average precision (mAP) of the YOLOv4 model for identifying white shrimp was 0.76. The accuracy of the appearing number is 0.85. The accuracy of dropping rate was 0.69. The accuracy of body length was 0.94. The body length changes linearly with time, and the coefficient of determination (R square) was 0.98, which is suitable for white shrimp over 3 cm. The height of the imaging system platform will affect the chance of shrimps appearing. The appearing number will obviously decrease and then increase after feeding. It shows the degree to which the shrimps are attracted away from the platform by the feeding, which is helpful for judging the appetite of the shrimps and adjusting the feeding amount. And the timing of feeding, the change trend of the dropping rate in the first half was similar to the appearing number, but it was not significant in the second half. The automatic feeder specially developed for waterway pools in this study can control feeding in the cloud, making night-time feeding feasible, and completed the automatic feeding experiment in Pingtung. The white shrimp IoT monitoring system proposed in this study overcomes the bottleneck of difficult observation of white shrimp in pond culture, provides important information for white shrimp growth and behavior patterns, and facilitates the automation of feeding control.